xref: /freebsd/sys/netinet/tcp_output.c (revision 09e8dea79366f1e5b3a73e8a271b26e4b6bf2e6a)
1 /*
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
34  * $FreeBSD$
35  */
36 
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 
53 #include <net/route.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #ifdef INET6
61 #include <netinet6/in6_pcb.h>
62 #include <netinet/ip6.h>
63 #include <netinet6/ip6_var.h>
64 #endif
65 #include <netinet/tcp.h>
66 #define	TCPOUTFLAGS
67 #include <netinet/tcp_fsm.h>
68 #include <netinet/tcp_seq.h>
69 #include <netinet/tcp_timer.h>
70 #include <netinet/tcp_var.h>
71 #include <netinet/tcpip.h>
72 #ifdef TCPDEBUG
73 #include <netinet/tcp_debug.h>
74 #endif
75 
76 #ifdef IPSEC
77 #include <netinet6/ipsec.h>
78 #endif /*IPSEC*/
79 
80 #include <machine/in_cksum.h>
81 
82 #ifdef notyet
83 extern struct mbuf *m_copypack();
84 #endif
85 
86 int path_mtu_discovery = 1;
87 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
88 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
89 
90 int ss_fltsz = 1;
91 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
92 	&ss_fltsz, 1, "Slow start flight size");
93 
94 int ss_fltsz_local = 4;
95 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
96 	&ss_fltsz_local, 1, "Slow start flight size for local networks");
97 
98 int     tcp_do_newreno = 1;
99 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno,
100         0, "Enable NewReno Algorithms");
101 /*
102  * Tcp output routine: figure out what should be sent and send it.
103  */
104 int
105 tcp_output(struct tcpcb *tp)
106 {
107 	struct socket *so = tp->t_inpcb->inp_socket;
108 	long len, win;
109 	int off, flags, error;
110 	struct mbuf *m;
111 	struct ip *ip = NULL;
112 	struct ipovly *ipov = NULL;
113 	struct tcphdr *th;
114 	u_char opt[TCP_MAXOLEN];
115 	unsigned ipoptlen, optlen, hdrlen;
116 	int idle, sendalot;
117 #if 0
118 	int maxburst = TCP_MAXBURST;
119 #endif
120 	struct rmxp_tao *taop;
121 	struct rmxp_tao tao_noncached;
122 #ifdef INET6
123 	struct ip6_hdr *ip6 = NULL;
124 	int isipv6;
125 
126 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
127 #endif
128 
129 	/*
130 	 * Determine length of data that should be transmitted,
131 	 * and flags that will be used.
132 	 * If there is some data or critical controls (SYN, RST)
133 	 * to send, then transmit; otherwise, investigate further.
134 	 */
135 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
136 	if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
137 		/*
138 		 * We have been idle for "a while" and no acks are
139 		 * expected to clock out any data we send --
140 		 * slow start to get ack "clock" running again.
141 		 *
142 		 * Set the slow-start flight size depending on whether
143 		 * this is a local network or not.
144 		 */
145 		int ss = ss_fltsz;
146 #ifdef INET6
147 		if (isipv6) {
148 			if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
149 				ss = ss_fltsz_local;
150 		} else
151 #endif /* INET6 */
152 		if (in_localaddr(tp->t_inpcb->inp_faddr))
153 			ss = ss_fltsz_local;
154 		tp->snd_cwnd = tp->t_maxseg * ss;
155 	}
156 	tp->t_flags &= ~TF_LASTIDLE;
157 	if (idle) {
158 		if (tp->t_flags & TF_MORETOCOME) {
159 			tp->t_flags |= TF_LASTIDLE;
160 			idle = 0;
161 		}
162 	}
163 again:
164 	sendalot = 0;
165 	off = tp->snd_nxt - tp->snd_una;
166 	win = min(tp->snd_wnd, tp->snd_cwnd);
167 
168 	flags = tcp_outflags[tp->t_state];
169 	/*
170 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
171 	 * state flags.
172 	 */
173 	if (tp->t_flags & TF_NEEDFIN)
174 		flags |= TH_FIN;
175 	if (tp->t_flags & TF_NEEDSYN)
176 		flags |= TH_SYN;
177 
178 	/*
179 	 * If in persist timeout with window of 0, send 1 byte.
180 	 * Otherwise, if window is small but nonzero
181 	 * and timer expired, we will send what we can
182 	 * and go to transmit state.
183 	 */
184 	if (tp->t_force) {
185 		if (win == 0) {
186 			/*
187 			 * If we still have some data to send, then
188 			 * clear the FIN bit.  Usually this would
189 			 * happen below when it realizes that we
190 			 * aren't sending all the data.  However,
191 			 * if we have exactly 1 byte of unsent data,
192 			 * then it won't clear the FIN bit below,
193 			 * and if we are in persist state, we wind
194 			 * up sending the packet without recording
195 			 * that we sent the FIN bit.
196 			 *
197 			 * We can't just blindly clear the FIN bit,
198 			 * because if we don't have any more data
199 			 * to send then the probe will be the FIN
200 			 * itself.
201 			 */
202 			if (off < so->so_snd.sb_cc)
203 				flags &= ~TH_FIN;
204 			win = 1;
205 		} else {
206 			callout_stop(tp->tt_persist);
207 			tp->t_rxtshift = 0;
208 		}
209 	}
210 
211 	len = (long)ulmin(so->so_snd.sb_cc, win) - off;
212 
213 	if ((taop = tcp_gettaocache(&tp->t_inpcb->inp_inc)) == NULL) {
214 		taop = &tao_noncached;
215 		bzero(taop, sizeof(*taop));
216 	}
217 
218 	/*
219 	 * Lop off SYN bit if it has already been sent.  However, if this
220 	 * is SYN-SENT state and if segment contains data and if we don't
221 	 * know that foreign host supports TAO, suppress sending segment.
222 	 */
223 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
224 		flags &= ~TH_SYN;
225 		off--, len++;
226 		if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
227 		    taop->tao_ccsent == 0)
228 			return 0;
229 	}
230 
231 	/*
232 	 * Be careful not to send data and/or FIN on SYN segments
233 	 * in cases when no CC option will be sent.
234 	 * This measure is needed to prevent interoperability problems
235 	 * with not fully conformant TCP implementations.
236 	 */
237 	if ((flags & TH_SYN) &&
238 	    ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
239 	     ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
240 		len = 0;
241 		flags &= ~TH_FIN;
242 	}
243 
244 	if (len < 0) {
245 		/*
246 		 * If FIN has been sent but not acked,
247 		 * but we haven't been called to retransmit,
248 		 * len will be -1.  Otherwise, window shrank
249 		 * after we sent into it.  If window shrank to 0,
250 		 * cancel pending retransmit, pull snd_nxt back
251 		 * to (closed) window, and set the persist timer
252 		 * if it isn't already going.  If the window didn't
253 		 * close completely, just wait for an ACK.
254 		 */
255 		len = 0;
256 		if (win == 0) {
257 			callout_stop(tp->tt_rexmt);
258 			tp->t_rxtshift = 0;
259 			tp->snd_nxt = tp->snd_una;
260 			if (!callout_active(tp->tt_persist))
261 				tcp_setpersist(tp);
262 		}
263 	}
264 	if (len > tp->t_maxseg) {
265 		len = tp->t_maxseg;
266 		sendalot = 1;
267 	}
268 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
269 		flags &= ~TH_FIN;
270 
271 	win = sbspace(&so->so_rcv);
272 
273 	/*
274 	 * Sender silly window avoidance.   We transmit under the following
275 	 * conditions when len is non-zero:
276 	 *
277 	 *	- We have a full segment
278 	 *	- This is the last buffer in a write()/send() and we are
279 	 *	  either idle or running NODELAY
280 	 *	- we've timed out (e.g. persist timer)
281 	 *	- we have more then 1/2 the maximum send window's worth of
282 	 *	  data (receiver may be limited the window size)
283 	 *	- we need to retransmit
284 	 */
285 	if (len) {
286 		if (len == tp->t_maxseg)
287 			goto send;
288 		/*
289 		 * NOTE! on localhost connections an 'ack' from the remote
290 		 * end may occur synchronously with the output and cause
291 		 * us to flush a buffer queued with moretocome.  XXX
292 		 *
293 		 * note: the len + off check is almost certainly unnecessary.
294 		 */
295 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
296 		    (idle || (tp->t_flags & TF_NODELAY)) &&
297 		    len + off >= so->so_snd.sb_cc &&
298 		    (tp->t_flags & TF_NOPUSH) == 0) {
299 			goto send;
300 		}
301 		if (tp->t_force)			/* typ. timeout case */
302 			goto send;
303 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
304 			goto send;
305 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
306 			goto send;
307 	}
308 
309 	/*
310 	 * Compare available window to amount of window
311 	 * known to peer (as advertised window less
312 	 * next expected input).  If the difference is at least two
313 	 * max size segments, or at least 50% of the maximum possible
314 	 * window, then want to send a window update to peer.
315 	 */
316 	if (win > 0) {
317 		/*
318 		 * "adv" is the amount we can increase the window,
319 		 * taking into account that we are limited by
320 		 * TCP_MAXWIN << tp->rcv_scale.
321 		 */
322 		long adv = min(win, (long)TCP_MAXWIN << tp->rcv_scale) -
323 			(tp->rcv_adv - tp->rcv_nxt);
324 
325 		if (adv >= (long) (2 * tp->t_maxseg))
326 			goto send;
327 		if (2 * adv >= (long) so->so_rcv.sb_hiwat)
328 			goto send;
329 	}
330 
331 	/*
332 	 * Send if we owe peer an ACK.
333 	 */
334 	if (tp->t_flags & TF_ACKNOW)
335 		goto send;
336 	if ((flags & TH_RST) ||
337 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
338 		goto send;
339 	if (SEQ_GT(tp->snd_up, tp->snd_una))
340 		goto send;
341 	/*
342 	 * If our state indicates that FIN should be sent
343 	 * and we have not yet done so, or we're retransmitting the FIN,
344 	 * then we need to send.
345 	 */
346 	if (flags & TH_FIN &&
347 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
348 		goto send;
349 
350 	/*
351 	 * TCP window updates are not reliable, rather a polling protocol
352 	 * using ``persist'' packets is used to insure receipt of window
353 	 * updates.  The three ``states'' for the output side are:
354 	 *	idle			not doing retransmits or persists
355 	 *	persisting		to move a small or zero window
356 	 *	(re)transmitting	and thereby not persisting
357 	 *
358 	 * callout_active(tp->tt_persist)
359 	 *	is true when we are in persist state.
360 	 * tp->t_force
361 	 *	is set when we are called to send a persist packet.
362 	 * callout_active(tp->tt_rexmt)
363 	 *	is set when we are retransmitting
364 	 * The output side is idle when both timers are zero.
365 	 *
366 	 * If send window is too small, there is data to transmit, and no
367 	 * retransmit or persist is pending, then go to persist state.
368 	 * If nothing happens soon, send when timer expires:
369 	 * if window is nonzero, transmit what we can,
370 	 * otherwise force out a byte.
371 	 */
372 	if (so->so_snd.sb_cc && !callout_active(tp->tt_rexmt) &&
373 	    !callout_active(tp->tt_persist)) {
374 		tp->t_rxtshift = 0;
375 		tcp_setpersist(tp);
376 	}
377 
378 	/*
379 	 * No reason to send a segment, just return.
380 	 */
381 	return (0);
382 
383 send:
384 	/*
385 	 * Before ESTABLISHED, force sending of initial options
386 	 * unless TCP set not to do any options.
387 	 * NOTE: we assume that the IP/TCP header plus TCP options
388 	 * always fit in a single mbuf, leaving room for a maximum
389 	 * link header, i.e.
390 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
391 	 */
392 	optlen = 0;
393 #ifdef INET6
394 	if (isipv6)
395 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
396 	else
397 #endif
398 	hdrlen = sizeof (struct tcpiphdr);
399 	if (flags & TH_SYN) {
400 		tp->snd_nxt = tp->iss;
401 		if ((tp->t_flags & TF_NOOPT) == 0) {
402 			u_short mss;
403 
404 			opt[0] = TCPOPT_MAXSEG;
405 			opt[1] = TCPOLEN_MAXSEG;
406 			mss = htons((u_short) tcp_mssopt(tp));
407 			(void)memcpy(opt + 2, &mss, sizeof(mss));
408 			optlen = TCPOLEN_MAXSEG;
409 
410 			if ((tp->t_flags & TF_REQ_SCALE) &&
411 			    ((flags & TH_ACK) == 0 ||
412 			    (tp->t_flags & TF_RCVD_SCALE))) {
413 				*((u_int32_t *)(opt + optlen)) = htonl(
414 					TCPOPT_NOP << 24 |
415 					TCPOPT_WINDOW << 16 |
416 					TCPOLEN_WINDOW << 8 |
417 					tp->request_r_scale);
418 				optlen += 4;
419 			}
420 		}
421  	}
422 
423  	/*
424 	 * Send a timestamp and echo-reply if this is a SYN and our side
425 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
426 	 * and our peer have sent timestamps in our SYN's.
427  	 */
428  	if ((tp->t_flags & (TF_REQ_TSTMP|TF_NOOPT)) == TF_REQ_TSTMP &&
429  	    (flags & TH_RST) == 0 &&
430 	    ((flags & TH_ACK) == 0 ||
431 	     (tp->t_flags & TF_RCVD_TSTMP))) {
432 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
433 
434  		/* Form timestamp option as shown in appendix A of RFC 1323. */
435  		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
436  		*lp++ = htonl(ticks);
437  		*lp   = htonl(tp->ts_recent);
438  		optlen += TCPOLEN_TSTAMP_APPA;
439  	}
440 
441  	/*
442 	 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
443 	 * options are allowed (!TF_NOOPT) and it's not a RST.
444  	 */
445  	if ((tp->t_flags & (TF_REQ_CC|TF_NOOPT)) == TF_REQ_CC &&
446  	     (flags & TH_RST) == 0) {
447 		switch (flags & (TH_SYN|TH_ACK)) {
448 		/*
449 		 * This is a normal ACK, send CC if we received CC before
450 		 * from our peer.
451 		 */
452 		case TH_ACK:
453 			if (!(tp->t_flags & TF_RCVD_CC))
454 				break;
455 			/*FALLTHROUGH*/
456 
457 		/*
458 		 * We can only get here in T/TCP's SYN_SENT* state, when
459 		 * we're a sending a non-SYN segment without waiting for
460 		 * the ACK of our SYN.  A check above assures that we only
461 		 * do this if our peer understands T/TCP.
462 		 */
463 		case 0:
464 			opt[optlen++] = TCPOPT_NOP;
465 			opt[optlen++] = TCPOPT_NOP;
466 			opt[optlen++] = TCPOPT_CC;
467 			opt[optlen++] = TCPOLEN_CC;
468 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
469 
470 			optlen += 4;
471 			break;
472 
473 		/*
474 		 * This is our initial SYN, check whether we have to use
475 		 * CC or CC.new.
476 		 */
477 		case TH_SYN:
478 			opt[optlen++] = TCPOPT_NOP;
479 			opt[optlen++] = TCPOPT_NOP;
480 			opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
481 						TCPOPT_CCNEW : TCPOPT_CC;
482 			opt[optlen++] = TCPOLEN_CC;
483 			*(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
484  			optlen += 4;
485 			break;
486 
487 		/*
488 		 * This is a SYN,ACK; send CC and CC.echo if we received
489 		 * CC from our peer.
490 		 */
491 		case (TH_SYN|TH_ACK):
492 			if (tp->t_flags & TF_RCVD_CC) {
493 				opt[optlen++] = TCPOPT_NOP;
494 				opt[optlen++] = TCPOPT_NOP;
495 				opt[optlen++] = TCPOPT_CC;
496 				opt[optlen++] = TCPOLEN_CC;
497 				*(u_int32_t *)&opt[optlen] =
498 					htonl(tp->cc_send);
499 				optlen += 4;
500 				opt[optlen++] = TCPOPT_NOP;
501 				opt[optlen++] = TCPOPT_NOP;
502 				opt[optlen++] = TCPOPT_CCECHO;
503 				opt[optlen++] = TCPOLEN_CC;
504 				*(u_int32_t *)&opt[optlen] =
505 					htonl(tp->cc_recv);
506 				optlen += 4;
507 			}
508 			break;
509 		}
510  	}
511 
512  	hdrlen += optlen;
513 
514 #ifdef INET6
515 	if (isipv6)
516 		ipoptlen = ip6_optlen(tp->t_inpcb);
517 	else
518 #endif
519 	if (tp->t_inpcb->inp_options)
520 		ipoptlen = tp->t_inpcb->inp_options->m_len -
521 				offsetof(struct ipoption, ipopt_list);
522 	else
523 		ipoptlen = 0;
524 #ifdef IPSEC
525 	ipoptlen += ipsec_hdrsiz_tcp(tp);
526 #endif
527 
528 	/*
529 	 * Adjust data length if insertion of options will
530 	 * bump the packet length beyond the t_maxopd length.
531 	 * Clear the FIN bit because we cut off the tail of
532 	 * the segment.
533 	 */
534 	if (len + optlen + ipoptlen > tp->t_maxopd) {
535 		/*
536 		 * If there is still more to send, don't close the connection.
537 		 */
538 		flags &= ~TH_FIN;
539 		len = tp->t_maxopd - optlen - ipoptlen;
540 		sendalot = 1;
541 	}
542 
543 /*#ifdef DIAGNOSTIC*/
544 #ifdef INET6
545  	if (max_linkhdr + hdrlen > MCLBYTES)
546 #else
547  	if (max_linkhdr + hdrlen > MHLEN)
548 #endif
549 		panic("tcphdr too big");
550 /*#endif*/
551 
552 	/*
553 	 * Grab a header mbuf, attaching a copy of data to
554 	 * be transmitted, and initialize the header from
555 	 * the template for sends on this connection.
556 	 */
557 	if (len) {
558 		if (tp->t_force && len == 1)
559 			tcpstat.tcps_sndprobe++;
560 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
561 			tcpstat.tcps_sndrexmitpack++;
562 			tcpstat.tcps_sndrexmitbyte += len;
563 		} else {
564 			tcpstat.tcps_sndpack++;
565 			tcpstat.tcps_sndbyte += len;
566 		}
567 #ifdef notyet
568 		if ((m = m_copypack(so->so_snd.sb_mb, off,
569 		    (int)len, max_linkhdr + hdrlen)) == 0) {
570 			error = ENOBUFS;
571 			goto out;
572 		}
573 		/*
574 		 * m_copypack left space for our hdr; use it.
575 		 */
576 		m->m_len += hdrlen;
577 		m->m_data -= hdrlen;
578 #else
579 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
580 		if (m == NULL) {
581 			error = ENOBUFS;
582 			goto out;
583 		}
584 #ifdef INET6
585 		if (MHLEN < hdrlen + max_linkhdr) {
586 			MCLGET(m, M_DONTWAIT);
587 			if ((m->m_flags & M_EXT) == 0) {
588 				m_freem(m);
589 				error = ENOBUFS;
590 				goto out;
591 			}
592 		}
593 #endif
594 		m->m_data += max_linkhdr;
595 		m->m_len = hdrlen;
596 		if (len <= MHLEN - hdrlen - max_linkhdr) {
597 			m_copydata(so->so_snd.sb_mb, off, (int) len,
598 			    mtod(m, caddr_t) + hdrlen);
599 			m->m_len += len;
600 		} else {
601 			m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len);
602 			if (m->m_next == 0) {
603 				(void) m_free(m);
604 				error = ENOBUFS;
605 				goto out;
606 			}
607 		}
608 #endif
609 		/*
610 		 * If we're sending everything we've got, set PUSH.
611 		 * (This will keep happy those implementations which only
612 		 * give data to the user when a buffer fills or
613 		 * a PUSH comes in.)
614 		 */
615 		if (off + len == so->so_snd.sb_cc)
616 			flags |= TH_PUSH;
617 	} else {
618 		if (tp->t_flags & TF_ACKNOW)
619 			tcpstat.tcps_sndacks++;
620 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
621 			tcpstat.tcps_sndctrl++;
622 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
623 			tcpstat.tcps_sndurg++;
624 		else
625 			tcpstat.tcps_sndwinup++;
626 
627 		MGETHDR(m, M_DONTWAIT, MT_HEADER);
628 		if (m == NULL) {
629 			error = ENOBUFS;
630 			goto out;
631 		}
632 #ifdef INET6
633 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
634 		    MHLEN >= hdrlen) {
635 			MH_ALIGN(m, hdrlen);
636 		} else
637 #endif
638 		m->m_data += max_linkhdr;
639 		m->m_len = hdrlen;
640 	}
641 	m->m_pkthdr.rcvif = (struct ifnet *)0;
642 #ifdef INET6
643 	if (isipv6) {
644 		ip6 = mtod(m, struct ip6_hdr *);
645 		th = (struct tcphdr *)(ip6 + 1);
646 		tcp_fillheaders(tp, ip6, th);
647 	} else
648 #endif /* INET6 */
649       {
650 	ip = mtod(m, struct ip *);
651 	ipov = (struct ipovly *)ip;
652 	th = (struct tcphdr *)(ip + 1);
653 	/* this picks up the pseudo header (w/o the length) */
654 	tcp_fillheaders(tp, ip, th);
655       }
656 
657 	/*
658 	 * Fill in fields, remembering maximum advertised
659 	 * window for use in delaying messages about window sizes.
660 	 * If resending a FIN, be sure not to use a new sequence number.
661 	 */
662 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
663 	    tp->snd_nxt == tp->snd_max)
664 		tp->snd_nxt--;
665 	/*
666 	 * If we are doing retransmissions, then snd_nxt will
667 	 * not reflect the first unsent octet.  For ACK only
668 	 * packets, we do not want the sequence number of the
669 	 * retransmitted packet, we want the sequence number
670 	 * of the next unsent octet.  So, if there is no data
671 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
672 	 * when filling in ti_seq.  But if we are in persist
673 	 * state, snd_max might reflect one byte beyond the
674 	 * right edge of the window, so use snd_nxt in that
675 	 * case, since we know we aren't doing a retransmission.
676 	 * (retransmit and persist are mutually exclusive...)
677 	 */
678 	if (len || (flags & (TH_SYN|TH_FIN))
679 	    || callout_active(tp->tt_persist))
680 		th->th_seq = htonl(tp->snd_nxt);
681 	else
682 		th->th_seq = htonl(tp->snd_max);
683 	th->th_ack = htonl(tp->rcv_nxt);
684 	if (optlen) {
685 		bcopy(opt, th + 1, optlen);
686 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
687 	}
688 	th->th_flags = flags;
689 	/*
690 	 * Calculate receive window.  Don't shrink window,
691 	 * but avoid silly window syndrome.
692 	 */
693 	if (win < (long)(so->so_rcv.sb_hiwat / 4) && win < (long)tp->t_maxseg)
694 		win = 0;
695 	if (win < (long)(tp->rcv_adv - tp->rcv_nxt))
696 		win = (long)(tp->rcv_adv - tp->rcv_nxt);
697 	if (win > (long)TCP_MAXWIN << tp->rcv_scale)
698 		win = (long)TCP_MAXWIN << tp->rcv_scale;
699 	th->th_win = htons((u_short) (win>>tp->rcv_scale));
700 
701 
702 	/*
703 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
704 	 * a 0 window.  This may cause the remote transmitter to stall.  This
705 	 * flag tells soreceive() to disable delayed acknowledgements when
706 	 * draining the buffer.  This can occur if the receiver is attempting
707 	 * to read more data then can be buffered prior to transmitting on
708 	 * the connection.
709 	 */
710 	if (win == 0)
711 		tp->t_flags |= TF_RXWIN0SENT;
712 	else
713 		tp->t_flags &= ~TF_RXWIN0SENT;
714 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
715 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
716 		th->th_flags |= TH_URG;
717 	} else
718 		/*
719 		 * If no urgent pointer to send, then we pull
720 		 * the urgent pointer to the left edge of the send window
721 		 * so that it doesn't drift into the send window on sequence
722 		 * number wraparound.
723 		 */
724 		tp->snd_up = tp->snd_una;		/* drag it along */
725 
726 	/*
727 	 * Put TCP length in extended header, and then
728 	 * checksum extended header and data.
729 	 */
730 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
731 #ifdef INET6
732 	if (isipv6)
733 		/*
734 		 * ip6_plen is not need to be filled now, and will be filled
735 		 * in ip6_output.
736 		 */
737 		th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
738 				       sizeof(struct tcphdr) + optlen + len);
739 	else
740 #endif /* INET6 */
741       {
742 	m->m_pkthdr.csum_flags = CSUM_TCP;
743 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
744 	if (len + optlen)
745 		th->th_sum = in_addword(th->th_sum,
746 		    htons((u_short)(optlen + len)));
747 
748 	/* IP version must be set here for ipv4/ipv6 checking later */
749 	KASSERT(ip->ip_v == IPVERSION,
750 	    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
751       }
752 
753 	/*
754 	 * In transmit state, time the transmission and arrange for
755 	 * the retransmit.  In persist state, just set snd_max.
756 	 */
757 	if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
758 		tcp_seq startseq = tp->snd_nxt;
759 
760 		/*
761 		 * Advance snd_nxt over sequence space of this segment.
762 		 */
763 		if (flags & (TH_SYN|TH_FIN)) {
764 			if (flags & TH_SYN)
765 				tp->snd_nxt++;
766 			if (flags & TH_FIN) {
767 				tp->snd_nxt++;
768 				tp->t_flags |= TF_SENTFIN;
769 			}
770 		}
771 		tp->snd_nxt += len;
772 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
773 			tp->snd_max = tp->snd_nxt;
774 			/*
775 			 * Time this transmission if not a retransmission and
776 			 * not currently timing anything.
777 			 */
778 			if (tp->t_rtttime == 0) {
779 				tp->t_rtttime = ticks;
780 				tp->t_rtseq = startseq;
781 				tcpstat.tcps_segstimed++;
782 			}
783 		}
784 
785 		/*
786 		 * Set retransmit timer if not currently set,
787 		 * and not doing an ack or a keep-alive probe.
788 		 * Initial value for retransmit timer is smoothed
789 		 * round-trip time + 2 * round-trip time variance.
790 		 * Initialize shift counter which is used for backoff
791 		 * of retransmit time.
792 		 */
793 		if (!callout_active(tp->tt_rexmt) &&
794 		    tp->snd_nxt != tp->snd_una) {
795 			if (callout_active(tp->tt_persist)) {
796 				callout_stop(tp->tt_persist);
797 				tp->t_rxtshift = 0;
798 			}
799 			callout_reset(tp->tt_rexmt, tp->t_rxtcur,
800 				      tcp_timer_rexmt, tp);
801 		}
802 	} else
803 		if (SEQ_GT(tp->snd_nxt + len, tp->snd_max))
804 			tp->snd_max = tp->snd_nxt + len;
805 
806 #ifdef TCPDEBUG
807 	/*
808 	 * Trace.
809 	 */
810 	if (so->so_options & SO_DEBUG)
811 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
812 #endif
813 
814 	/*
815 	 * Fill in IP length and desired time to live and
816 	 * send to IP level.  There should be a better way
817 	 * to handle ttl and tos; we could keep them in
818 	 * the template, but need a way to checksum without them.
819 	 */
820 	/*
821 	 * m->m_pkthdr.len should have been set before cksum calcuration,
822 	 * because in6_cksum() need it.
823 	 */
824 #ifdef INET6
825 	if (isipv6) {
826 		/*
827 		 * we separately set hoplimit for every segment, since the
828 		 * user might want to change the value via setsockopt.
829 		 * Also, desired default hop limit might be changed via
830 		 * Neighbor Discovery.
831 		 */
832 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb,
833 					       tp->t_inpcb->in6p_route.ro_rt ?
834 					       tp->t_inpcb->in6p_route.ro_rt->rt_ifp
835 					       : NULL);
836 
837 		/* TODO: IPv6 IP6TOS_ECT bit on */
838 #ifdef IPSEC
839 		if (ipsec_setsocket(m, so) != 0) {
840 			m_freem(m);
841 			error = ENOBUFS;
842 			goto out;
843 		}
844 #endif /*IPSEC*/
845 		error = ip6_output(m,
846 			    tp->t_inpcb->in6p_outputopts,
847 			    &tp->t_inpcb->in6p_route,
848 			    (so->so_options & SO_DONTROUTE), NULL, NULL);
849 	} else
850 #endif /* INET6 */
851     {
852 	struct rtentry *rt;
853 	ip->ip_len = m->m_pkthdr.len;
854 #ifdef INET6
855  	if (INP_CHECK_SOCKAF(so, AF_INET6))
856  		ip->ip_ttl = in6_selecthlim(tp->t_inpcb,
857  					    tp->t_inpcb->in6p_route.ro_rt ?
858  					    tp->t_inpcb->in6p_route.ro_rt->rt_ifp
859  					    : NULL);
860  	else
861 #endif /* INET6 */
862 	ip->ip_ttl = tp->t_inpcb->inp_ip_ttl;	/* XXX */
863 	ip->ip_tos = tp->t_inpcb->inp_ip_tos;	/* XXX */
864 	/*
865 	 * See if we should do MTU discovery.  We do it only if the following
866 	 * are true:
867 	 *	1) we have a valid route to the destination
868 	 *	2) the MTU is not locked (if it is, then discovery has been
869 	 *	   disabled)
870 	 */
871 	if (path_mtu_discovery
872 	    && (rt = tp->t_inpcb->inp_route.ro_rt)
873 	    && rt->rt_flags & RTF_UP
874 	    && !(rt->rt_rmx.rmx_locks & RTV_MTU)) {
875 		ip->ip_off |= IP_DF;
876 	}
877 #ifdef IPSEC
878  	ipsec_setsocket(m, so);
879 #endif /*IPSEC*/
880 	error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
881 	    (so->so_options & SO_DONTROUTE), 0);
882     }
883 	if (error) {
884 
885 		/*
886 		 * We know that the packet was lost, so back out the
887 		 * sequence number advance, if any.
888 		 */
889 		if (tp->t_force == 0 || !callout_active(tp->tt_persist)) {
890 			/*
891 			 * No need to check for TH_FIN here because
892 			 * the TF_SENTFIN flag handles that case.
893 			 */
894 			if ((flags & TH_SYN) == 0)
895 				tp->snd_nxt -= len;
896 		}
897 
898 out:
899 		if (error == ENOBUFS) {
900 	                if (!callout_active(tp->tt_rexmt) &&
901                             !callout_active(tp->tt_persist))
902 	                        callout_reset(tp->tt_rexmt, tp->t_rxtcur,
903                                       tcp_timer_rexmt, tp);
904 			tcp_quench(tp->t_inpcb, 0);
905 			return (0);
906 		}
907 		if (error == EMSGSIZE) {
908 			/*
909 			 * ip_output() will have already fixed the route
910 			 * for us.  tcp_mtudisc() will, as its last action,
911 			 * initiate retransmission, so it is important to
912 			 * not do so here.
913 			 */
914 			tcp_mtudisc(tp->t_inpcb, 0);
915 			return 0;
916 		}
917 		if ((error == EHOSTUNREACH || error == ENETDOWN)
918 		    && TCPS_HAVERCVDSYN(tp->t_state)) {
919 			tp->t_softerror = error;
920 			return (0);
921 		}
922 		return (error);
923 	}
924 	tcpstat.tcps_sndtotal++;
925 
926 	/*
927 	 * Data sent (as far as we can tell).
928 	 * If this advertises a larger window than any other segment,
929 	 * then remember the size of the advertised window.
930 	 * Any pending ACK has now been sent.
931 	 */
932 	if (win > 0 && SEQ_GT(tp->rcv_nxt+win, tp->rcv_adv))
933 		tp->rcv_adv = tp->rcv_nxt + win;
934 	tp->last_ack_sent = tp->rcv_nxt;
935 	tp->t_flags &= ~TF_ACKNOW;
936 	if (tcp_delack_enabled)
937 		callout_stop(tp->tt_delack);
938 #if 0
939 	/*
940 	 * This completely breaks TCP if newreno is turned on.  What happens
941 	 * is that if delayed-acks are turned on on the receiver, this code
942 	 * on the transmitter effectively destroys the TCP window, forcing
943 	 * it to four packets (1.5Kx4 = 6K window).
944 	 */
945 	if (sendalot && (!tcp_do_newreno || --maxburst))
946 		goto again;
947 #endif
948 	if (sendalot)
949 		goto again;
950 	return (0);
951 }
952 
953 void
954 tcp_setpersist(tp)
955 	register struct tcpcb *tp;
956 {
957 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
958 	int tt;
959 
960 	if (callout_active(tp->tt_rexmt))
961 		panic("tcp_setpersist: retransmit pending");
962 	/*
963 	 * Start/restart persistance timer.
964 	 */
965 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
966 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
967 	callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
968 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
969 		tp->t_rxtshift++;
970 }
971